Videos

Core Ideas in Artificial Intelligence (AI): From Perceptron to Transformers

This lecture argues that the most important foundational concept in computer science is the abstraction of the natural neuron that underlies modern AI. Dr. Tizhoosh traces the evolution from the perceptron to contemporary deep learning architectures such as Transformers, highlighting how this core idea has shaped the understanding and development of intelligent systems.

Foundation Models and Information Retrieval in Digital Pathology

This talk explains how foundation models — deep neural networks trained on massive datasets — can handle a wide variety of tasks. It also highlights the benefits of retrieval-augmented generation, particularly in domains such as pathology, where combining evidence retrieval with generation improves reliability.

Image Search: Past, Present and the Path Forward

The lecture traces three decades of research on image search in histopathology, moving from handcrafted features and barcodes to deep learning and divide-conquer-combine methods. It highlights challenges such as indexing whole-slide images, patch-based retrieval, storage and validation. Despite significant technical progress with systems such as Yottixel, clinical deployment remains limited. The talk emphasizes the need for robust validation and practical integration to truly impact pathology practice.

Learning or Searching: Foundations Models and Information Retrieval in Digital Pathology

This talk highlights the problem of intraobserver and interobserver variability in pathology, showing how inconsistent diagnoses threaten patient care. The speaker contrasts two AI approaches:

  • Large classification models. This approach aims for high accuracy through classification but risk-limited generalization and hallucination.
  • Search and retrieval systems. This approach grounds decisions in evidence from past cases.

Dr. Tizhoosh argues that reducing variability should be the primary goal of AI in medicine and that retrieval-augmented methods may offer more trustworthy, evidence-based support for clinical consensus than classification alone.